Building Humanoids For Real Homes
We deploy humanoid robots today using human-in-the-loop telepresence. Every interaction trains the next generation of physical AI.
To build AI for homes, our robots go to homes – not labs, not factory floors, not laundromats, not coffeeshops.
The Technical Thesis
Full autonomy in unstructured home environments is years away – for everyone. The manipulation, reasoning, and safety requirements are unsolved problems. Many humanoid companies show visions of home robots, but build applications for commercial spaces.
We took a different path: deploy to homes now with human operators, generate real-world data at scale, and use that data to progressively automate. This isn’t a workaround – it’s the fastest route to physical AI that works.
Every teleoperation session creates training episodes. Every home deployment creates massive data variability. The result is a compounding advantage: more deployments → more data → better autonomy → more deployments.
Robody — The Hardware
Designed for homes, proven in the field
- Height: 1.65 m
- Weight: 60 kg
- Arms: 6 DoF, 1.5 kg payload each
- Neck: 3 DoF
- End-effector: 2-finger gripper (finray-based proprietary design)
- Actuation: QDDs + Dynamixel integrated actuators
- Mobility: Differential drive, stable platform
- Battery: 6h runtime, self-dock capability, inductive charging
- Sensors: 4K fisheye RGB, mm-wave radar, stereo microphone
- Compute: Nvidia Jetson Orin + Orin Nano
- Middleware: ROS2
Safety isn’t a spec we optimized for – it’s the starting point.
The robot is intentionally weak: 1.5kg per arm won’t win any lifting competitions, but it’s enough to pour tea, carry a plate, or hand someone their medication—and gentle enough that physical contact with a person is never a risk.
In real apartments, you’re always navigating around something, so the robot inevitably bumps into tabletops, doorframes, cabinets. That’s why the whole body is wrapped in a soft skin that absorbs impacts, fully enclosed with no exposed joints or mechanisms where a curious hand could get pinched. The mobile base is slim enough for narrow hallways, stable enough that it won’t tip over.
And because Robody will live in someone’s home for years, it should feel like it belongs there. It wears clothes—families dress it how they like. And it has a face. Not a screen, not a blinking light, but a face that moves, expresses, and makes eye contact. In any interaction with a person, that changes everything.
You cannot be at two places at once
Robody’s telepresence system is built around embodiment, not remote control. The operator experiences a true first-person perspective: the VR headset streams stereoscopic video directly from Robody’s fisheye camera array, spatialized audio is captured through on-board microphones, and every hand motion is mapped one-to-one onto the robot’s end-effector. The result is a control loop that feels less like “operating a machine” and more like inhabiting a body.
Under the hood, the full telepresence stack runs on WebRTC with state-of-the-art end-to-end encryption, optimized for low-latency bidirectional streaming. Typical glass-to-glass latency stays under 200 ms on standard residential internet, enabling natural motion and conversational flow. Motion commands, multi-camera video, audio, and auxiliary telemetry all synchronize through our custom real-time transport layer.
For human expressiveness, the VR headset captures 60+ facial blend shapes in real time. These signals drive a photorealistic operator avatar rendered via Gaussian splatting, which is displayed on Robody’s face. The remote user doesn’t interact with a flat screen or stylized cartoon—they interact with you.
The interface is intentionally minimal: most operators reach stable control after just a few minutes. Across hundreds of users—care professionals, researchers, family members, government officials—training curves are consistently steep. Expert-level proficiency typically emerges within a few hours.
What We’re Building Next
- Action transformers for bi-manual mobile manipulation in homes
- Sample-efficient Learning from Demonstration at scale
- Safe physical interaction with vulnerable populations
- Network-adaptive shared autonomy
- Ultra-low-latency telepresence over unreliable networks
- Fleet intelligence
If you want to work on:
- Real-world robotics deployment (not just simulation)
- Embodied AI with immediate impact
- Full-stack robotics challenges
- A team that’s been doing this for a decade
WANT TO WORK WITH US?
Real-world robotics deployment (not just simulation), Embodied AI with immediate impact, Full-stack robotics challenges and a team that’s been doing this for a decade sound great? Don’t wait! Apply: apply@devanthro.com